On Fri, Mar 03, 2017 at 06:45:40PM +0200, Laurent Pinchart wrote:
> - I haven't seen any proposal how a heap-based solution could be used in a
> generic distribution. This needs to be figured out before committing to any
> API/ABI.
Two replies from my side:
- Just because a patch doesn't solve world hunger isn't really a good
reason to reject it.
- Heap doesn't mean its not resizeable (but I'm not sure that's really
your concern).
- Imo ION is very much part of the picture here to solve this for real. We
need to bits:
* Be able to allocate memory from specific pools, not going through a
specific driver. ION gives us that interface. This is e.g. also needed
for "special" memory, like SMA tries to expose.
* Some way to figure out how&where to allocate the buffer object. This
is purely a userspace problem, and this is the part the unix memory
allocator tries to solve. There's no plans in there for big kernel
changes, instead userspace does a dance to reconcile all the
constraints, and one of the constraints might be "you have to allocate
this from this special ION heap". The only thing the kernel needs to
expose is which devices use which ION heaps (we kinda do that
already), and maybe some hints of how they can be generalized (but I
guess stuff like "minimal pagesize of x KB" is also fulfilled by any
CMA heap is knowledge userspace needs).
Again I think waiting for this to be fully implemented before we merge any
part is going to just kill any upstreaming efforts. ION in itself, without
the full buffer negotiation dance seems clearly useful (also for stuff
like SMA), and having it merged will help with moving the buffer
allocation dance forward.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
